GPU Buying Guide

The chipset brand is the company that designs the actual graphics processor at the heart of the card. Everything else on the listing — the cooler, the lights, the box art — is built around this one decision, so start here.

• Nvidia — The market leader and the safe default. Nvidia holds a commanding lead in ray tracing performance and its DLSS upscaling is the most mature and most widely supported. It also dominates in professional and AI workloads thanks to CUDA, which many creative and machine learning applications are built specifically to use. The trade-off is price: you generally pay a premium for the same raw rasterization performance.
• AMD — The value play. AMD Radeon cards typically offer more memory and more traditional gaming performance per dollar than an equivalently priced Nvidia card. FSR upscaling has closed much of the quality gap, and recent generations have improved ray tracing considerably. Software support in niche creative apps is thinner, and driver reputation, while much improved, still makes some buyers hesitate.
• Intel — The disruptor. Intel Arc cards are aggressively priced at the budget and lower-midrange end and include genuinely good media encoding and respectable ray tracing for the money. The catch is that performance can be inconsistent in older games built on legacy APIs, and the driver track record is the newest of the three. Also verify your motherboard supports Resizable BAR, as Arc cards lose significant performance without it.

If you want maximum performance, ray tracing, or CUDA-dependent software, Nvidia is the answer. If you want the most frames per dollar in ordinary gaming, look hard at AMD. If you're building a budget machine on a modern platform, Intel deserves a serious look.

The model number is the single most important number on the listing — it determines roughly how fast the card is. Both Nvidia and AMD use a tiering system where the first digits indicate generation and the later digits indicate performance tier within that generation.

• Flagship tier (RTX 5090, RTX 4090, RX 7900 XTX) — Uncompromised 4K gaming at high refresh rates, heavy 3D rendering, and local AI work. These cards are enormous, draw enormous power, and cost more than many complete PCs. Only worth it if you have a 4K high-refresh display and a power supply to match.
• High-end tier (RTX 5080, RTX 5070 Ti, RX 9070 XT, RX 7900 XT) — The sweet spot for 1440p at very high frame rates and competent 4K gaming. Strong ray tracing, plenty of memory, and far more sensible pricing than flagship parts.
• Midrange tier (RTX 5070, RTX 5060 Ti, RTX 4070, RX 9070, RX 9060 XT, RX 7800 XT, Arc B580) — Where most buyers should be shopping. Excellent 1440p performance, playable 4K in less demanding titles, and reasonable power draw. This is the best value band in the entire market.
• Entry tier (RTX 5060, RTX 5050, RTX 4060, RX 7600, RX 6600, Arc B570, Arc A750) — Solid 1080p gaming at high settings and serviceable 1440p if you're willing to turn a few sliders down. Perfect for esports titles and modest budgets.
• Basic and display-output tier (GTX 1650, GT 1030, GT 710, RX 550, RX 6400) — These are not gaming cards in any modern sense. Buy them to add display outputs to an office machine or to get video out of a CPU without integrated graphics. Avoid them if you plan to play anything released this decade.
• Workstation tier (RTX A4000, RTX A2000) — Professional cards with certified drivers for CAD, simulation, and engineering software. They trade raw gaming speed for stability, compact form factors, and validated application support. Buy these only if your software vendor specifically recommends them.

Match the tier to your monitor, not to your ego. A midrange card paired with a 1440p display will feel far better than a flagship card bottlenecked by an old 1080p 60Hz panel.

Architecture is the generational engineering behind the chip. It determines which features the card supports, how efficiently it uses power, and how long it will keep receiving meaningful driver and feature updates.

• Blackwell — Nvidia's current generation, powering the RTX 50 series. Brings GDDR7 memory, the newest DLSS feature set including multi frame generation, and substantially improved AI acceleration hardware.
• Ada Lovelace — Nvidia's RTX 40 series. Still excellent, extremely power efficient, and supports frame generation. Now widely available at discounted prices, which makes it one of the best value targets on the used and clearance market.
• Ampere — Nvidia's RTX 30 series. Capable rasterization performance but weaker ray tracing and no frame generation support. Only compelling at a steep discount.
• Turing — Nvidia's RTX 20 and GTX 16 series. First-generation ray tracing that is largely too slow to use. Consider these entry-level cards today.
• Pascal and Maxwell — GTX 10 series and older. These are legacy parts with no ray tracing, no DLSS, and declining driver support. Buy only for basic display duties.
• RDNA 4 — AMD's current generation behind the RX 9000 series. A major leap in ray tracing performance and the architecture required for FSR 4 upscaling.
• RDNA 3 — AMD's RX 7000 series. Strong traditional gaming performance with generous memory, though ray tracing lags. Frequently discounted and often a smart buy.
• RDNA 2 — AMD's RX 6000 series. Aging but still perfectly playable at 1080p and 1440p, and often available at very aggressive prices.
• Polaris — The RX 500 series. Genuinely obsolete for modern gaming and frequently sold as heavily used mining hardware. Approach with caution.
• Xe2-HPG — Intel's second-generation Arc B-series architecture. Much better driver maturity and efficiency than the first attempt.
• Xe-HPG — Intel's first Arc A-series generation. Good value but the most inconsistent performance of any current architecture.

Newer architectures don't just run faster — they unlock software features that older cards can never receive. If two cards perform similarly, always take the newer architecture.

Video memory, or VRAM, holds textures, frame buffers, and assets the GPU is actively working with. Running out of it causes stuttering and sudden frame rate collapse that no amount of raw GPU power can fix.

• 1GB, 2GB, and 3GB — Legacy capacities suitable only for desktop use, video playback, and very old games. Do not buy these for gaming.
• 4GB — Genuinely insufficient for modern titles even at 1080p. Many current games will refuse to load high-quality textures at this capacity.
• 6GB — The bare minimum for 1080p gaming today, and already showing strain in newer releases. Acceptable only on a tight budget with expectations set accordingly.
• 8GB — The most common capacity in the entry and lower-midrange market. Fine for 1080p at high settings, but increasingly limiting at 1440p with ray tracing enabled.
• 10GB, 11GB, and 12GB — A comfortable band for 1440p gaming with headroom for high-resolution textures. 12GB is currently the practical sweet spot for most buyers.
• 16GB — Ideal for 4K gaming, heavy modding, video editing, and running local AI models. This is the capacity to target if you want the card to stay relevant for years.
• 20GB and 24GB — Enthusiast and creator territory. Necessary for large 3D scenes, high-resolution video timelines, and sizable AI workloads.
• 32GB and 48GB — Professional and flagship capacities aimed at rendering farms, machine learning, and specialized workstation tasks. Massive overkill for gaming alone.

VRAM is the spec most likely to make a card feel obsolete before its raw performance does. When choosing between two similar cards, the one with more memory will almost always age better.

Memory type determines how quickly the card can move data in and out of its VRAM. Two cards with identical memory capacity can perform very differently if one uses a faster memory standard.

• GDDR7 — The newest and fastest standard, found on current-generation flagship and high-end cards. Delivers substantially more bandwidth per pin, which particularly helps at 4K and with ray tracing.
• GDDR6X — A high-bandwidth variant used on Nvidia's upper-tier previous-generation cards. Very fast, though it runs noticeably warm.
• GDDR6 — The workhorse standard across the majority of modern cards from all three brands. Reliable, efficient, and more than adequate for mainstream gaming.
• GDDR5X and GDDR5 — Previous-generation standards found on older cards. Functional but significantly slower, and a clear sign you're looking at aging hardware.
• DDR4 and DDR3 — System memory repurposed for graphics use on the cheapest possible cards. Dramatically slower than any GDDR standard and a major red flag on a gaming card listing.
• HBM2 — High Bandwidth Memory, a stacked design offering enormous bandwidth. Rare and expensive, appearing mainly on older enthusiast and professional cards.

Watch for budget cards that advertise a large memory number while quietly using DDR4 — the capacity is meaningless if the bandwidth can't feed the GPU.

Bus width is the size of the pipe connecting the GPU to its memory, measured in bits. Combined with memory type, it determines total memory bandwidth, and it's a spec manufacturers often quietly narrow on budget cards.

• 64-Bit — Extremely narrow. Found on the lowest-end cards and a strong indicator of severe bandwidth limitations, especially at higher resolutions.
• 96-Bit — Still very restrictive. Acceptable only for basic 1080p and light workloads.
• 128-Bit — Standard for entry and lower-midrange cards. Adequate for 1080p, though it can bottleneck the card at 1440p and above unless paired with large cache.
• 192-Bit — Comfortable midrange width. A good match for 1440p gaming and typically paired with 12GB of memory.
• 256-Bit — High-end territory offering excellent bandwidth for 1440p and 4K. Usually paired with 16GB configurations.
• 320-Bit and 384-Bit — Enthusiast widths reserved for flagship cards driving 4K and beyond. These pipes are wide enough that memory rarely becomes the limiting factor.
• 512-Bit — Ultra-flagship and professional class. Enormous bandwidth for the most demanding rendering and AI workloads.

Modern cards partly compensate for narrow buses with large on-chip cache, so don't judge on bus width alone — but a 64-Bit or 96-Bit bus on a card marketed for gaming is still worth questioning.

This spec tells you the resolution the card is realistically built to drive in games, not merely the maximum it can output to a display. Buying a card matched to your monitor is the single most effective way to avoid overspending or underperforming.

• 1080p — Full HD, still the most popular gaming resolution. Entry-level cards handle it comfortably at high settings, and the frame rate ceiling is generous enough for competitive play.
• 1440p — The current enthusiast sweet spot, offering a visible sharpness upgrade over 1080p without the brutal performance cost of 4K. Requires a midrange card at minimum.
• 4K — Four times the pixels of 1080p and a punishing workload. Realistically needs a high-end or flagship card, and upscaling technologies become essentially mandatory in demanding titles.
• 8K — Largely a marketing checkbox for gaming. Useful for professional video work and large-format displays, but no current card plays modern games well at native 8K.

Buy for the monitor you own or are about to buy. A card rated well above your display's resolution mostly delivers frames you'll never see.

Boost clock is the peak frequency the GPU core reaches under load. It's a useful tiebreaker between cards using the same chip, but a poor way to compare cards across different architectures or brands.

• Under 1500 MHz — Typical of older generations and low-power professional cards. Not inherently bad, but usually indicates aging silicon.
• 1500 MHz to 2000 MHz — Common across previous-generation midrange cards and many current entry-level parts.
• 2000 MHz to 2500 MHz — The standard operating range for most current-generation cards from every brand.
• 2500 MHz to 3000 MHz — Found on high-end and factory-overclocked models with strong cooling. Requires good case airflow to sustain.
• Over 3000 MHz — Flagship and heavily overclocked territory. Impressive on paper, but real-world gains over a standard model are typically in the low single-digit percentages.

Never compare clock speeds across different architectures — a 2000 MHz current-generation card will crush a 2500 MHz card from three generations ago. Use this spec only when choosing between versions of the same model.

The interface is the PCIe slot standard the card uses to talk to your motherboard. It's backward and forward compatible, so a mismatch won't stop the card from working, but it can cost you performance in specific cases.

• PCIe 5.0 — The current standard on the newest cards and motherboards. Offers bandwidth well beyond what today's games require, so the practical benefit is mostly future-proofing.
• PCIe 4.0 — The mainstream standard and more than sufficient for essentially every consumer card on the market.
• PCIe 3.0 — Still perfectly usable for most cards. The exception is budget cards that use only four or eight PCIe lanes, which can lose meaningful performance on a 3.0 board.
• PCIe 2.0 — Legacy. Only a concern if you're upgrading a very old system, where it may genuinely limit a modern card.

Interface generation rarely matters for full-width cards, but if you're pairing a budget card with an older motherboard, check the card's lane count as well as its PCIe generation.

The power connector tells you what your power supply needs to provide. Getting this wrong means the card either won't boot or will crash under load, so verify it against your PSU before you order.

• Slot Powered — Draws all power from the motherboard slot, up to 75 watts. No cables needed, making these ideal for prebuilt systems and small form factor machines with weak power supplies.
• 6-Pin — Adds 75 watts. Found on entry-level and older cards with modest power draw.
• 8-Pin — Adds 150 watts and is the most common single connector on midrange cards.
• Dual 8-Pin — Standard for high-end cards. Make sure your PSU has two genuinely separate cables rather than one cable with a daisy-chained splitter.
• Triple 8-Pin — Enthusiast and heavily overclocked cards. Demands a high-wattage, high-quality power supply.
• 12VHPWR — The 16-pin high-power connector introduced with recent Nvidia cards, capable of delivering up to 600 watts through one cable. Early revisions had well-documented issues with melting when not fully seated, so push it in until it clicks.
• 12V-2x6 — The revised version of the same connector with improved contact design and better safety margins. Functionally identical in use, and the version you want on any new build.

Check both your PSU's wattage and its available connectors before buying. Adapters work in a pinch, but a power supply with native cables is always the safer and cleaner solution.

Cooling determines how loud your system is and whether the card can sustain its advertised clock speeds. A well-cooled card runs faster for longer, so this is not a spec to dismiss as cosmetic.

• Fanless — Passive heatsink only, completely silent. Limited to very low-power cards, and ideal for home theater PCs and quiet office machines.
• Single Fan — Compact and suited to small form factor builds. Effective on low and entry-level cards but tends to get loud when pushed.
• Dual Fan — The mainstream standard. A good balance of cooling capacity, noise, and physical size for most midrange cards.
• Triple Fan — Standard on high-end cards. Excellent thermal performance and low noise, but these cards are long and heavy, so measure your case first.
• Quad Fan — Rare, oversized designs on extreme flagship models. Superb cooling at the cost of enormous physical dimensions.
• Blower — Uses a single centrifugal fan to exhaust hot air directly out the back of the case. Louder than open-air designs but invaluable in cramped cases or multi-GPU workstation setups where recirculating hot air is a problem.
• Liquid Cooled — An integrated closed-loop cooler with a radiator. Outstanding temperatures and very quiet operation, but requires radiator mounting space and adds cost.
• Hybrid Cooled — Combines a liquid loop for the GPU core with fans for the memory and power delivery. The best thermal performance available, at the highest price and complexity.

More fans generally means quieter, cooler operation — but also a bigger card. Confirm your case's maximum GPU length and slot clearance before you commit.

Slot width describes how many expansion slots the card physically occupies in your case. This is a pure compatibility check, and it's the most common reason a new card ends up going back in the box.

• Single Slot — Occupies one slot. Found on low-power and professional cards, and essential for dense multi-card workstation builds.
• Dual Slot — The traditional standard and still the most common. Fits comfortably in nearly every case.
• 2.5 Slot — Slightly oversized to accommodate a thicker heatsink. Usually fine, but it will block the adjacent slot's usability.
• Triple Slot — Common on modern high-end cards with large coolers. Requires you to plan around losing three slots of internal space.
• 3.5 Slot and Quad Slot — Extreme flagship designs. These are genuinely massive and will not fit in many mid-tower cases, let alone anything smaller.

Measure your case's internal clearance in three dimensions — length, height, and thickness — before ordering. Slot width is where big cards most often collide with side panels, cables, and drive cages.

Video outputs determine what monitors you can connect and at what resolution and refresh rate. The port version matters as much as the port type, since an older standard can cap your display's capabilities.

• HDMI 2.1 — Supports 4K at high refresh rates and 8K output. The port you want for modern TVs and high-refresh 4K monitors.
• HDMI 2.0 — Handles 4K at 60Hz. Fine for older displays but limiting on anything faster.
• Mini HDMI — A compact version found on small form factor and low-profile cards. Requires an adapter or a specific cable.
• DisplayPort 2.1 — The highest-bandwidth consumer standard available, built for very high resolution and refresh rate combinations. The best choice for cutting-edge gaming monitors.
• DisplayPort 1.4 — The long-standing standard, capable of 4K at high refresh rates with compression. Still excellent for the vast majority of monitors.
• Mini DisplayPort — A smaller connector common on professional and compact cards. Functionally identical to full-size DisplayPort with an adapter.
• USB-C — Carries video alongside data and power. Useful for VR headsets, portable monitors, and single-cable desk setups.
• DVI — A legacy digital connector. Only relevant if you're keeping an older monitor that lacks anything newer.
• VGA — Analog and thoroughly obsolete. Its presence on a listing is a reliable sign of very old hardware.

Count the ports as well as reading their versions — if you run three monitors, confirm the card has three outputs you can actually use simultaneously.

Features cover both the software capabilities baked into the chip and the physical design touches added by the board partner. The software features affect performance directly; the hardware ones affect longevity, noise, and how the card looks in your case.

• Ray Tracing — Hardware-accelerated realistic lighting, reflections, and shadows. Visually transformative in supported games but very demanding, so pair it with upscaling on anything but a high-end card.
• DLSS 4 — Nvidia's newest upscaling and frame generation suite, exclusive to current-generation hardware. Delivers major frame rate gains with excellent image quality.
• DLSS 3 — The previous version, including single frame generation. Still highly effective and supported across a wide library of games.
• FSR 4 — AMD's latest upscaler, using machine learning for a significant image quality jump. Requires current-generation Radeon hardware.
• FSR 3 — AMD's broadly compatible upscaling and frame generation solution. Works across a wide range of hardware including competitors' cards, making it a useful fallback.
• XeSS — Intel's upscaling technology. Performs best on Arc hardware but runs on other vendors' cards too.
• Frame Generation — Inserts AI-generated frames between rendered ones to boost apparent smoothness. Excellent for single-player games, less suited to competitive titles where input latency matters most.
• Overclocked — A factory tune above reference clocks. Real gains are typically modest, so only pay a small premium for it.
• RGB Lighting — Purely cosmetic. Check whether it syncs with your motherboard software if you care about a coordinated build.
• Zero RPM Fan — Fans stop entirely under light load, making the card silent when browsing or idle. A genuinely valuable quality-of-life feature.
• Dual BIOS — A physical switch toggling between quiet and performance fan profiles, and a useful recovery option if a BIOS update fails.
• Metal Backplate — Adds rigidity and helps dissipate heat from the rear of the board. Worth having on any large card.
• Vapor Chamber — A premium heat-spreading technology that moves heat away from the core more evenly than heat pipes alone. Common on flagship coolers.
• Copper Heat Pipe — The standard heat transfer method in modern coolers. More pipes and thicker pipes generally mean better thermals.
• Anti-Sag Bracket — A support strut preventing heavy cards from drooping in the slot over time. Increasingly necessary as cards get larger, and nice to get in the box rather than buying separately.

Upscaling support has become as important as raw horsepower — a card with a strong upscaler often outperforms a nominally faster card without one. Among physical features, zero RPM fans and a metal backplate deliver the most real-world benefit.

The right card depends entirely on what you'll actually do with it. A card that's perfect for video editing may be a poor value for competitive gaming, and vice versa.

• Gaming — Prioritize raw performance for your target resolution, plus upscaling support and enough VRAM to avoid texture stutter. Match the tier to your monitor rather than buying the most expensive option available.
• Video Editing — Focus on VRAM capacity and hardware encoding support. Nvidia's NVENC encoder and Intel's media engine are both excellent, and 12GB or more makes timeline scrubbing on high-resolution footage far smoother.
• 3D Rendering — VRAM is king here, since complex scenes must fit entirely in memory. Nvidia holds a strong advantage due to CUDA and OptiX support in most major renderers.
• AI and Machine Learning — VRAM capacity determines what model sizes you can run at all. CUDA support makes Nvidia the practical default, and 16GB should be considered the entry point for serious local work.
• Workstation — Certified drivers, stability under sustained load, and validated application support matter more than peak frame rates. Check what your specific software vendor recommends.
• Streaming — Hardware encoders let you broadcast without taxing the CPU or tanking your in-game performance. Any recent Nvidia, AMD, or Intel card handles this competently.
• Multi-Monitor — Count the outputs and confirm the card supports your combined resolution and refresh rate across all displays simultaneously.
• Home Theater — Prioritize silence, small size, and HDMI 2.1 with modern codec support. A fanless or single-fan low-profile card is usually ideal.
• Office — Basic display output and low power draw are all you need. A slot-powered card keeps the build simple and quiet.
• Mining — Largely obsolete as a consumer pursuit. Its main relevance today is as a warning: be cautious of heavily discounted used cards that may have run under sustained load for years.

Be honest about your primary workload before you shop. Buying a creator-focused card for pure gaming, or a gaming card for professional certified software, is the fastest way to overpay for capability you'll never use.

Board partners take the chipset and build the actual card around it — the cooler, the power delivery, the warranty, and the support. The chip determines performance; the partner determines noise levels, temperatures, build quality, and what happens if something goes wrong.

• Asus — A premium partner known for the ROG Strix and TUF Gaming lines. Excellent cooling and build quality across the board, with pricing to match. Their midrange TUF cards are often the best-cooled option in their price bracket.
• MSI — Strong, consistent designs spanning the Ventus, Gaming X, and Suprim lines. Ventus targets value, Gaming X hits the sweet spot for cooling and noise, and Suprim competes at the premium end.
• Gigabyte — Broad availability and competitive pricing, with the Windforce, Gaming OC, and Aorus tiers. Generally solid value, though their entry models can be louder than rivals.
• Zotac — Nvidia-only, frequently among the most affordable options, and notable for genuinely compact models that fit small form factor builds where other cards won't.
• Sapphire — An AMD specialist with an outstanding reputation. The Nitro+ and Pulse lines are consistently regarded as among the best-built Radeon cards available.
• PowerColor — Another AMD-focused partner offering strong value through the Hellhound and Red Devil lines, with particularly good acoustic tuning.
• XFX — AMD-exclusive with distinctive cooler designs and competitive pricing. Their Merc and Swift lines cool well and are generally quiet.
• PNY — Nvidia-focused with a significant presence in professional and workstation cards. Consumer models are typically no-frills and priced accordingly.
• ASRock — Primarily AMD, offering aggressive pricing and increasingly competent cooler designs. A good value pick if you find a deal.
• Palit, Gainward, and Inno3D — Budget-oriented Nvidia partners. Performance is identical to pricier cards using the same chip, but coolers are simpler and warranty support can be less convenient depending on your region.
• Acer — A newer entrant with the Predator and Nitro lines, notable as one of the more visible partners for Intel Arc cards.

The chipset sets your performance ceiling, so don't overpay for a premium partner card unless you specifically want better cooling, lower noise, or a stronger warranty. Whichever brand you choose, always check current customer reviews for the exact model — quality varies between lines within the same brand.